Interpretable Machine Learning for Screening Liquid Cooling Materials via Enthalpy of Vaporization

Liquid cooling (LC) is increasingly important for thermal management in high-power electronic systems, yet the identification of suitable working fluids remains limited by an insufficient understanding of molecular structure-property relationships and reliance on empirical screening. Herein, we develop an interpretable machine learning (ML) framework for the systematic evaluation of LC materials based on the enthalpy of vaporization, a key thermodynamic parameter governing phase-change heat transfer capacity. A curated data set of 2474 organic compounds was constructed using SMILES-derived molecular descriptors, and multiple regression models were trained and optimized. The gradient boosting decision tree (GBDT) model achieved a test-set R2 of 0.99 under random splitting and was subsequently applied to screen 136 candidate compounds. Additional assessments of synthetic accessibility, safety, and operating temperature compatibility enabled the prioritization of promising LC materials. Interpretability analyses identify four key molecular features, polarity, conformational flexibility, intermolecular interactions, and molecular size, as influential factors in vaporization energetics within the studied chemical space. Complementary symbolic regression and qualitative molecular dynamics simulations provide mechanistic insight into the role of molecular flexibility. This work establishes a reproducible and interpretable computational strategy for data-informed screening and rational design of LC materials, with potential extension to other thermophysical properties and energy-relevant liquid systems.